Forecasting Inflation Using Many Predictors

This thesis analyzes the performance of linear and non-linear dimensionality reduction techniques in the context of inflation forecasting. In particular, Principal Component Analysis (PCA) and three Partial Least Squares (PLS) variants are used as representative of linear methods. Non-linear models include Squared Principal Components (SQPC), Kernel PCA, and Kernel PLS. The findings indicate that factor models substantially improve the forecasting performance in comparison to univariate autoregressive models. More importantly, although this paper finds non-linear models to dominate the linear ones in specific subsamples and forecast horizons, it concludes that there is rather limited room for improvement in the forecasting methodology concerning non-linear models. A key reason is that these models come with the major drawback that their performance is highly dependent on the choice of hyperparameters.